Friday, 4 September 2026
Artificial intelligence is beginning to change asset and wealth management at a deeper level than simply helping analysts read documents faster. Investment firms are now exploring how AI can combine research, market information, portfolio risk, client data and institutional knowledge into continuously operating intelligence systems, potentially changing how investment decisions are researched, communicated and ultimately implemented. That transition was the focus of the AI4 2026 panel “Alpha to Automation: How AI Is Reshaping Asset Management,” moderated by John Divine, Assistant Managing Editor for Investing at U.S. News & World Report, with specialists from Vanguard, State Street and JPMorgan Chase alongside expertise in agentic AI and financial data science.
The central message was that the industry is moving beyond the first generation of AI applications. Summarising an earnings call, searching regulatory filings or preparing research more quickly can already provide useful productivity improvements. The larger opportunity comes when those capabilities are connected with portfolio exposures, market developments, company announcements, risk information and proprietary institutional knowledge. Instead of providing an analyst with a faster research tool, such a system could create an intelligence layer around the entire investment process, allowing an analyst examining a company to receive not only a summary of its latest earnings announcement but also changes in market conditions, relevant portfolio exposures, external events and potential risks.
One area attracting particular attention is earnings analysis. Corporate earnings calls contain large amounts of information that can affect company valuations, but interpreting them involves considerably more than extracting financial figures. Executives choose their language carefully, meaning tone, confidence, changes in wording and what management avoids discussing can sometimes be as significant as the numbers themselves. The panel discussed research into agentic systems capable of examining earnings-call transcripts and producing assessments across different time horizons, potentially ranging from the following day to several weeks.
Such systems could combine the transcript with sentiment, market information and external developments in an attempt to identify signals that would be difficult for an analyst to assemble manually at comparable speed. This does not mean AI has discovered a reliable method of predicting share prices. The panellists repeatedly emphasised the limitations involved. Financial markets are influenced by enormous numbers of interconnected factors, while large language models can hallucinate information and remain unreliable when handling some numerical tasks.
The potential investment advantage may instead come from AI’s ability to examine nonlinear relationships across very large amounts of information. One example discussed was identifying whether markets have overreacted or underreacted to new information. Automated trading systems already respond rapidly to news, but those responses can themselves create opportunities if the resulting price movement becomes disconnected from the broader information available. AI could potentially compare the initial market reaction with company fundamentals, historical behaviour, external events and other signals to identify such anomalies, although whether this consistently generates investment alpha remains an open research question rather than an established outcome.
The discussion also highlighted an important distinction between large language models and the broader field of artificial intelligence. Financial institutions have used statistical models, machine learning, anomaly detection and quantitative systems for years. The emerging opportunity is increasingly about combining these established technologies with generative AI and agentic systems rather than expecting one large language model to perform every function.
That distinction becomes especially important where calculations are involved. A large language model might be useful for extracting information from documents, interpreting language or identifying relevant data, but deterministic software can still perform the underlying financial calculation. A discounted cash-flow model, for example, does not need a language model to calculate its mathematics. This type of architecture could become important to financial institutions seeking greater control over AI, with language models handling uncertain information while conventional software performs calculations and applies fixed rules.
Human judgement is consequently unlikely to disappear from asset management in the near future. The panel generally saw some of the greatest immediate AI opportunities in research and investment ideation, where professionals must process large quantities of information. AI can examine filings, earnings calls and other datasets and potentially identify patterns or questions that an analyst might investigate further. Portfolio construction and client communication require a different level of caution because investment decisions depend not only on available information but also on risk tolerance, objectives, personal circumstances and professional judgement.
This reflects a wider principle emerging across financial AI: the closer an automated system gets to moving money or making fiduciary decisions, the greater the requirement for controls, explainability and human accountability. For established financial institutions, however, one of the biggest obstacles may not be the models themselves. It is the operating structure surrounding them.
Many financial organisations still rely on processes developed decades ago. A client request might move through several departments, with each team adding information, conducting checks or approving part of the transaction. Introducing AI separately into each stage might make every department somewhat faster without addressing whether the original process is still necessary. Brinda Menon of JPMorgan Chase argued that the more transformative question is whether the process should be redesigned entirely. Rather than asking how AI can make stages A, B, C and D more efficient, companies can ask how they can move directly from A to D and whether some of the intermediate stages remain necessary.
That difference separates automation from transformation. Using AI to produce an existing report more quickly improves productivity. Redesigning the organisation so the report is no longer required changes the operating model. The same principle applies to investment research. Giving analysts faster access to information is useful, but connecting research, market developments, portfolio risks and institutional knowledge into a common intelligence environment could have a substantially greater effect.
Institutional knowledge may ultimately become one of the industry’s most valuable AI assets. Investment firms contain decades of expertise distributed between databases, research documents, policies, previous transactions and the experience of employees. Much of the most valuable knowledge remains inside people’s heads. If organisations can capture more of that expertise and make it accessible through AI, they could create an institutional intelligence system that survives individual employees and becomes available across the organisation.
This could be particularly significant as experienced professionals retire. Instead of losing part of the firm’s knowledge whenever a senior analyst, portfolio manager or adviser leaves, AI systems could potentially preserve elements of their processes, reasoning and accumulated expertise. Wealth management presents another major opportunity, particularly where AI could make financial guidance more accessible to people who currently have little or no access to professional advice.
State Street’s Sheema Osmani described work aimed at developing trusted AI-supported financial guidance and expanding access to financial knowledge. The underlying argument is that large numbers of people make important financial decisions without having access to a professional adviser. AI could potentially narrow that gap by providing personalised guidance at a scale impossible through traditional adviser-only models.
Personalisation in wealth management is more complicated than assigning investors to conventional demographic categories. People’s financial priorities change throughout their lives. Marriage, children, business ownership, retirement, inheritance, philanthropy and changes in wealth can substantially alter what an investor needs. AI potentially allows personalisation to become continuous rather than periodic, adapting as circumstances, priorities and behaviour change.
This does not necessarily diminish the importance of financial advisers. The more likely model is a division of responsibilities in which AI handles more information processing while advisers concentrate on the parts of wealth management that depend most heavily on trust, judgement and understanding individual circumstances.
Cost is another increasingly important issue as financial institutions move from experimental AI projects into production. The panel suggested that simply measuring the cost of individual tokens can provide a misleading picture of AI economics. A cheap model that produces unreliable results and requires extensive human checking may ultimately cost more than an expensive model that completes the task correctly. A more meaningful measure could therefore be the cost per successfully completed task.
That becomes particularly important with agentic AI. A single business process may involve several specialised agents communicating with one another, retrieving external information, checking outputs and applying guardrails. Every additional step consumes computing resources. Financial institutions are therefore experimenting with model routing, where relatively simple tasks are assigned to smaller and cheaper models while expensive frontier models are reserved for problems requiring more sophisticated reasoning.
The economics could improve as inference technology becomes more efficient, but the panel stressed that the industry remains too early in its development to know precisely where long-term AI operating costs will settle. The more important economic question may ultimately be the value produced. If AI simply reduces the cost of an existing process by a few percentage points, computing expenditure will remain closely scrutinised. If it allows an investment firm to redesign a process, substantially increase productivity or serve a much larger client base, the economics look very different.
Risk remains the counterweight to that opportunity. Financial institutions cannot simply accept answers from systems they do not understand, particularly where investment recommendations or client assets are involved. Black-box models, inconsistent responses and hallucinations create obvious concerns in a regulated industry.
The panel outlined several approaches to reducing that risk. AI can be required to show the sources used in reaching a conclusion, identify assumptions and provide confidence measures. Systems can also be designed to fail safely when information is insufficient rather than generating an answer regardless. Another approach is deliberately restricting what the AI controls. Language models can retrieve and interpret information while deterministic systems handle calculations and rule-based decisions. Agent activity can also be traced, providing a record of what information was searched, how it was processed and what actions were proposed.
Testing will become equally important. AI systems cannot be evaluated only against situations similar to those on which they were developed. The State Street discussion described using a substantial proportion of out-of-distribution examples when evaluating systems so that unexpected situations form part of testing rather than appearing only after deployment. This is particularly important in investment management because the events that cause the greatest losses are often precisely those that differ from normal market conditions.
For smaller investment firms without the resources of the world’s largest financial institutions, the panel’s recommendation was comparatively straightforward: begin with a narrow problem that is understood well. Document extraction, information retrieval or another repetitive process can provide an initial application where outputs are relatively easy to evaluate. Organisations can then reuse what they learn about models, data, controls and governance as they move towards more sophisticated applications.
The objective should not be to attach AI to every existing process. It should be to identify where intelligence and automation genuinely change the economics or quality of the work. The panel also challenged one of the more exaggerated claims surrounding AI and investment management: that autonomous agents will simply solve financial markets and generate predictable profits.
Financial markets are adaptive systems populated by investors reacting to one another. AI agents can exhibit biases, make incorrect assumptions and respond unpredictably just as human investors can. Asset management combines mathematics with judgement, behaviour and uncertainty, making it fundamentally different from a problem with a single correct solution.
That may ultimately determine how AI reshapes the industry. The near-term winner is unlikely to be an autonomous investment machine that replaces analysts, portfolio managers and advisers. It is more likely to be the investment organisation that learns how to combine machines capable of processing enormous amounts of information with humans capable of questioning what those machines conclude.
The transition from alpha to automation is therefore not simply about automating investment management. It is about rebuilding the investment operating model around a new division of labour between data, algorithms, AI agents and human judgement.
Source: CIJ.World Research & Analysis Team
